Agentic web

The agentic web is the ordinary web plus a layer of conventions that let software agents — rather than people driving browsers — discover what a site offers, prove who they are, learn what they may do, read its content, invoke its capabilities, coordinate with other agents, pay for what they use, present results to a person, and be held to account afterwards. None of this replaces HTTP, HTML or DNS. It sits on top of them: a set of well-known files, endpoints, headers and cards that a site can add so an agent does not have to guess. This site maps that layer as nine steps and treats every claim about it as something to source or mark as unverified, not something to assume.

What "agentic" adds

The conventional web was built for a person with a browser: pages carry visual layout, forms expect a human to fill them in, and a click is the unit of action. An agent arriving at the same URL has none of a browser's affordances to fall back on unless it renders the page like one, which is slow, brittle, and easy to break with a redesign. What "agentic" adds is a parallel, machine-native surface: a /.well-known/agent-card.json that states an A2A endpoint and its skills instead of a "Contact us" button; an MCP endpoint that exposes tools with typed schemas instead of a search box; a robots.txt with separate crawler tokens for search inclusion, model training and user-triggered fetches instead of one blanket disallow. The word does not describe a new network or a new browser. It describes a site choosing to publish, in a form an agent can parse without inference, the same things it already publishes for people.

What it is not

Agentic AI is a capability: a model that plans across multiple steps and calls tools to get something done. It can exist entirely inside one vendor's walls — a coding assistant editing files, a customer-support bot querying an internal database — with no reference to any other site's protocols at all. The agentic web is the opposite half of the picture: it is what a site publishes so that somebody else's agent, which it does not control and may never have heard of, can find and use it correctly. An agent can be highly capable and interact with zero agentic-web infrastructure; a site can publish a flawless Agent Card and still be talked to by a very simple bot.

The Semantic Web proposed a universal, machine-readable ontology of meaning — RDF triples, OWL classes, a web where any two systems could in principle reason about the same facts. The agentic web makes no such bet. Its protocols are imperative and narrow: fetch this card, call this endpoint, sign this request, settle this payment. There is no shared ontology requirement and no expectation that an agent "understands" a site beyond what a card or schema tells it to do next.

Web3 is a project about decentralised ownership and settlement, chiefly through blockchains. The agentic web borrows blockchain rails where they are useful for machine-to-machine payment — x402 and several of its adjacent proposals settle in stablecoins, and identity proposals such as ERC-8004 anchor agent registries on-chain — but most of the agentic web's standards layer (discovery files, agent cards, robots directives, HTTP signatures) has nothing to do with decentralisation and runs on ordinary client-server infrastructure that any single company can operate alone.

Browser automation is an agent pretending to be a person: it opens a real browser, reads the rendered DOM or accessibility tree, and clicks buttons and fills forms designed for human hands. It works against any site, including ones that have done nothing to accommodate agents, but it is slow, breaks when a layout changes, and gives the site no way to distinguish a considerate agent from an adversarial scraper. The agentic web is the alternative a site can offer instead: a native interface — an MCP tool, an A2A skill, a signed request — that does not require pretending to be a browser at all. WebMCP sits at the boundary between the two: it lets a page register tools that an agent embedded in the visitor's own browser can call directly, rather than having that agent click through the rendered UI, but as of Chrome's origin-trial rollout announced 9 June 2026 it is still experimental and origin-scoped, not a universally available browser capability.

The nine steps

#	Step	Question	Classic-web analog	Standards
1	Discover	What is here, and where?	sitemap, .well-known	12
2	Identify	Who is this agent? Who is this site?	TLS certificate, WHOIS	10
3	Permit	What may the agent do here?	robots.txt, terms of use	15
4	Read	Give me the content	HTML, RSS	5
5	Act	Do something for me	forms, REST APIs	12
6	Coordinate	Talk to another agent	none; agent-to-agent is new	10
7	Pay	Settle the bill	HTTP 402, finally used	12
8	Present	Show a human the result	the browser	3
9	Account	Prove what happened; be judged	logs, audit trails	3

An agent's path across a site can be described independently of any one protocol: it has to discover what is there, identify itself and the site, learn what it is permitted to do, read content, invoke a capability, coordinate with other agents where more than one is involved, pay if the resource is not free, present the outcome to a person, and leave a record that can be checked afterwards. This site organises every standard it tracks under one of these nine steps; each has a classic-web analogue that most of the concepts descend from.

How the pieces fit together

The steps are not independent of each other in practice. Discovery and identity typically arrive together: an A2A client fetches a site's Agent Card at /.well-known/agent-card.json (itself a use of the well-known-URI convention that RFC 8615 reserved in May 2019) and the card's own content — supported authentication schemes, provider identity — is what lets the client decide whether to trust what it just found. Permission sits alongside both: a robots.txt with distinct tokens for search crawling, training and user-triggered fetches (OpenAI's OAI-SearchBot, GPTBot and ChatGPT-User are documented as three separate policies, and Anthropic and Perplexity publish an equivalent three-way split) tells an agent what it may do with what it has found before it acts on it.

Invocation then depends on what discovery handed over: an MCP client that has parsed a server's tools uses that same session to call them; an A2A client that has read a skill's input/output modes structures its request accordingly. Payment, where it appears, is a further step layered on top of invocation rather than a replacement for it — Google's Agent Payments Protocol (AP2) chains an intent mandate, a cart mandate and a payment mandate as cryptographically signed, scoped authorisations so a person can let an agent transact without handing it standing account access, while x402 instead resumes an HTTP request that received a 402 status by attaching a signed payment payload, optionally verified and settled by a third-party facilitator rather than the resource server itself. Both approaches assume identity and permission were already established upstream; neither one does discovery's job for it.

What is shipping today

Coverage varies sharply by step. The Model Context Protocol, published by Anthropic and now governed under the Agentic AI Foundation and the Linux Foundation, is implemented by multiple independent clients — Claude Desktop, Cursor and Zed among them — against its current 2026-07-28 specification, and is the protocol behind the OpenAI Apps SDK's tool ecosystem. The Agent2Agent Protocol (A2A), originated by Google and now under Linux Foundation governance, reports a stable 1.0.x line with adoption named by Google's own Agent Development Kit, by Microsoft's Agent Framework (from 28 April 2026) and by Microsoft Work IQ (from 21 August 2026). x402, from Coinbase and now stewarded by an x402 Foundation, has multiple independent facilitator services running in production across several blockchain networks handling verification and settlement on a resource server's behalf, alongside servers that "self-facilitate" the same logic in-process. Web Bot Auth's signed-request approach to bot identification has separate documented implementations from Google, Cloudflare and Vercel. AP2 launched with more than sixty named partner organisations. Against that, most of the other protocols this site tracks — the majority of the discovery and coordinate rows in particular — have a specification and, at most, one implementing organisation; this site marks that distinction on every protocol page rather than treating "has a GitHub repository" as evidence of use.

Timeline, 2024–2026

The agentic web's standards did not arrive at once. Jeremy Howard's llms.txt proposal, a plain-Markdown convention for pointing language models at a site's documentation, dates to 2024, with no formal governance behind it. Anthropic published the Model Context Protocol in November 2024. Google followed with the Agent2Agent Protocol in April 2025, now on Linux Foundation track. Coinbase's x402 followed in May 2025, later gaining its own foundation and a 2026 v2 revision. IBM's Agent Communication Protocol, a separate agent-to-agent effort from its BeeAI team, was archived as a standalone specification on 29 August 2025, with its own repository stating it had been folded into A2A. Google published its Agent Payments Protocol (AP2) on 16 September 2025 with FIDO Alliance governance and more than sixty launch partners. Google and industry partners announced Agentic Resource Discovery (ARD) on 17 June 2026, with the specification reaching its v0.91 text on 26 August 2026. The Model Context Protocol's most recent dated specification revision is 2026-07-28. This site's own probe of ARD adoption across fifty domains ran on 2026-09-04.

Colliding acronyms

"ACP" is not one protocol; it is three, from three unrelated organisations, doing three different jobs at three different steps of this map. Zed's Agent Client Protocol connects code editors to coding agents, at the Act step. IBM's Agent Communication Protocol was an agent-to-agent messaging protocol at the Coordinate step, archived in August 2025 and, by its own maintainers' account, merged into A2A. OpenAI and Stripe's Agentic Commerce Protocol lets an agent complete a checkout with a merchant, at the Pay step, and is unrelated to either of the other two. Published guides and summaries routinely state a fact about one of these as though it applied to all three; the full picture, including a fourth candidate that could not be confirmed as a distinct specification, is on the ACP term page.

Frequently confused claims

Three claims recur in vendor and secondary material about this space and do not hold up as stated. Google-Extended does not control Google Search or AI Overviews. It is a separate token from Googlebot, documented as governing use of crawled content by Gemini apps and Vertex AI; Search and AI Overviews are covered by the ordinary Googlebot crawl, so disallowing Google-Extended alone does not remove a site from Google's AI-powered search surfaces. No mainstream answer engine documents reading llms.txt. Google's own generative-search guidance states plainly that no special markup or llms.txt file is required for its AI features, and no first-party documentation from ChatGPT Search, Perplexity, Claude, or Bing/Copilot lists it as something they fetch or use as a ranking or citation input; the file is a real, dated 2024 proposal with real adoption as a navigation aid, but that is a different claim from being consumed by answer engines. "ACP merged into A2A" is true only of IBM's Agent Communication Protocol. It says nothing about Zed's Agent Client Protocol or about the OpenAI/Stripe Agentic Commerce Protocol, both of which remain independent, active specifications; treating the merger as if it settled anything about "ACP" in general conflates three unrelated projects that happen to share an initialism.